Automated Poker Card Detection System
The client needed to identify and classify poker cards in online screenshots with over 95% accuracy to process large volumes of hands. The challenge included variability across 16 visual styles, occlusions in 9-player tables, resolutions from 720p to 4K, compression degradation and confusion between similar symbols (8/B, Q/K) that generated >15% errors with generic models. Additionally, the system needed to work in real-time (>15 FPS) on standard hardware and automatically adapt to new platforms without retraining.
Description
Computer Vision System for Poker Analysis
We developed an advanced poker card detection and classification system that uses artificial intelligence to analyze online game screenshots in real-time. Combining Faster R-CNN for detection and ResNet/YOLO for classification, this system achieves 98.3% accuracy in card identification under variable visual conditions. Designed to operate at 19.2 FPS on standard hardware, the system automatically adapts to different poker platforms, visual styles and lighting conditions, revolutionizing automatic game analysis for professional players and online poker platform operators.
Technologies
Objectives
Automatically detect the position of all cards in an online poker screenshot with >95% accuracy under variable conditions
Correctly classify the value and suit of each card with >95% accuracy, even on visually similar cards
Process images in real-time (minimum 15 FPS) for live game analysis and decision making
Create an adaptive system that works on various interface designs and graphic styles of different poker platforms without needing specific retraining
Generate an extensive and diverse training dataset for model robustness with over 8,000 labeled images
Implement a rigorous cross-validation system to properly evaluate performance on new platforms not seen during training
Optimize architecture for use with standard consumer hardware, enabling deployment on mid-range equipment
Project Gallery
Bayesian Marketing Mix Modeling for Global Fashion Retailer
The client faced significant inefficiencies in allocating their annual marketing budget of US$ 3MM distributed across 17 different channels. Their last-click attribution system over-estimated lower-funnel channels by up to 250%, generating approximately US$ 700K annual waste on saturated Display. Additionally, they lacked visibility on adstock effects and channel interactions, resulting in decisions based more on intuition than causal data. They needed to reduce daily prediction error (MAPE) below 15% and create a budget optimizer that could run in less than 90 minutes for weekly planning.
Flight Delay Prediction using Advanced Machine Learning
The airline faced significant losses estimated at US$ 40MM annually due to flight delays, generating additional operational costs, passenger compensation and reputational damage. Traditional prediction methods showed poor performance (F1-score < 0.55), and did not adequately capture route-specific, time-specific or seasonal delay patterns. They also lacked capacity to identify anomalies or cluster delay types to implement differentiated strategies. The goal was to reduce delay-associated costs by at least 5% through accurate predictions that would allow preventive measures to be taken with sufficient advance notice.
Hydroponic Crop Optimization with AI
High resource consumption and variable yield in hydroponic systems without ability to predict problems
Ready to Transform Your Business with AI?
Book a demo today and discover how our AI solutions can drive growth and efficiency for your organization.